Computing linkage disequilibrium aware genome embeddings using autoencoders.

MOTIVATION: The completion of the genome has paved the way for genome-wide association studies (GWAS), which explained certain proportions of heritability. GWAS are not optimally suited to detect non-linear effects in disease risk, possibly hidden in non-additive interactions (epistasis). Alternative methods for epistasis detection using, e.g. deep neural networks (DNNs) are currently under active development. However, DNNs are constrained by finite computational resources, which can be rapidly depleted due to increasing complexity with the sheer size of the genome. Besides, the curse of dimensionality complicates the task of capturing meaningful genetic patterns for DNNs; therefore necessitates dimensionality reduction. RESULTS: We propose a method to compress single nucleotide polymorphism (SNP) data, while leveraging the linkage disequilibrium (LD) structure and preserving potential epistasis. This method involves clustering correlated SNPs into haplotype blocks and training per-block autoencoders to learn a compressed representation of the block's genetic content. We provide an adjustable autoencoder design to accommodate diverse blocks and bypass extensive hyperparameter tuning. We applied this method to genotyping data from Project MinE, and achieved 99% average test reconstruction accuracy-i.e. minimal information loss-while compressing the input to nearly 10% of the original size. We demonstrate that haplotype-block based autoencoders outperform linear Principal Component Analysis (PCA) by approximately 3% chromosome-wide accuracy of reconstructed variants. To the extent of our knowledge, our approach is the first to simultaneously leverage haplotype structure and DNNs for dimensionality reduction of genetic data. AVAILABILITY AND IMPLEMENTATION: Data are available for academic use through Project MinE at https://www.projectmine.com/research/data-sharing/, contingent upon terms and requirements specified by the source studies. Code is available at https://github.com/gizem-tas/haploblock-autoencoders.

StrainXpress: strain aware metagenome assembly from short reads.

Next-generation sequencing-based metagenomics has enabled to identify microorganisms in characteristic habitats without the need for lengthy cultivation. Importantly, clinically relevant phenomena such as resistance to medication, virulence or interactions with the environment can vary already within species. Therefore, a major current challenge is to reconstruct individual genomes from the sequencing reads at the level of strains, and not just the level of species. However, strains of one species can differ only by minor amounts of variants, which makes it difficult to distinguish them. Despite considerable recent progress, related approaches have remained fragmentary so far. Here, we present StrainXpress, as a comprehensive solution to the problem of strain aware metagenome assembly from next-generation sequencing reads. In experiments, StrainXpress reconstructs strain-specific genomes from metagenomes that involve up to >1000 strains and proves to successfully deal with poorly covered strains. The amount of reconstructed strain-specific sequence exceeds that of the current state-of-the-art approaches by on average 26.75% across all data sets (first quartile: 18.51%, median: 26.60%, third quartile: 35.05%).

OGRE: Overlap Graph-based metagenomic Read clustEring.

MOTIVATION: The microbes that live in an environment can be identified from the combined genomic material, also referred to as the metagenome. Sequencing a metagenome can result in large volumes of sequencing reads. A promising approach to reduce the size of metagenomic datasets is by clustering reads into groups based on their overlaps. Clustering reads are valuable to facilitate downstream analyses, including computationally intensive strain-aware assembly. As current read clustering approaches cannot handle the large datasets arising from high-throughput metagenome sequencing, a novel read clustering approach is needed. In this article, we propose OGRE, an Overlap Graph-based Read clustEring procedure for high-throughput sequencing data, with a focus on shotgun metagenomes. RESULTS: We show that for small datasets OGRE outperforms other read binners in terms of the number of species included in a cluster, also referred to as cluster purity, and the fraction of all reads that is placed in one of the clusters. Furthermore, OGRE is able to process metagenomic datasets that are too large for other read binners into clusters with high cluster purity. CONCLUSION: OGRE is the only method that can successfully cluster reads in species-specific clusters for large metagenomic datasets without running into computation time- or memory issues. AVAILABILITYAND IMPLEMENTATION: Code is made available on Github (https://github.com/Marleen1/OGRE). SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

De novo assembly of viral quasispecies using overlap graphs.

A viral quasispecies, the ensemble of viral strains populating an infected person, can be highly diverse. For optimal assessment of virulence, pathogenesis, and therapy selection, determining the haplotypes of the individual strains can play a key role. As many viruses are subject to high mutation and recombination rates, high-quality reference genomes are often not available at the time of a new disease outbreak. We present SAVAGE, a computational tool for reconstructing individual haplotypes of intra-host virus strains without the need for a high-quality reference genome. SAVAGE makes use of either FM-index-based data structures or ad hoc consensus reference sequence for constructing overlap graphs from patient sample data. In this overlap graph, nodes represent reads and/or contigs, while edges reflect that two reads/contigs, based on sound statistical considerations, represent identical haplotypic sequence. Following an iterative scheme, a new overlap assembly algorithm that is based on the enumeration of statistically well-calibrated groups of reads/contigs then efficiently reconstructs the individual haplotypes from this overlap graph. In benchmark experiments on simulated and on real deep-coverage data, SAVAGE drastically outperforms generic de novo assemblers as well as the only specialized de novo viral quasispecies assembler available so far. When run on ad hoc consensus reference sequence, SAVAGE performs very favorably in comparison with state-of-the-art reference genome-guided tools. We also apply SAVAGE on two deep-coverage samples of patients infected by the Zika and the hepatitis C virus, respectively, which sheds light on the genetic structures of the respective viral quasispecies.

Overlap graph-based generation of haplotigs for diploids and polyploids.

MOTIVATION: Haplotype-aware genome assembly plays an important role in genetics, medicine and various other disciplines, yet generation of haplotype-resolved de novo assemblies remains a major challenge. Beyond distinguishing between errors and true sequential variants, one needs to assign the true variants to the different genome copies. Recent work has pointed out that the enormous quantities of traditional NGS read data have been greatly underexploited in terms of haplotig computation so far, which reflects that methodology for reference independent haplotig computation has not yet reached maturity. RESULTS: We present POLYploid genome fitTEr (POLYTE) as a new approach to de novo generation of haplotigs for diploid and polyploid genomes of known ploidy. Our method follows an iterative scheme where in each iteration reads or contigs are joined, based on their interplay in terms of an underlying haplotype-aware overlap graph. Along the iterations, contigs grow while preserving their haplotype identity. Benchmarking experiments on both real and simulated data demonstrate that POLYTE establishes new standards in terms of error-free reconstruction of haplotype-specific sequence. As a consequence, POLYTE outperforms state-of-the-art approaches in various relevant aspects, where advantages become particularly distinct in polyploid settings. AVAILABILITY AND IMPLEMENTATION: POLYTE is freely available as part of the HaploConduct package at https://github.com/HaploConduct/HaploConduct, implemented in Python and C++. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Full-length de novo viral quasispecies assembly through variation graph construction.

MOTIVATION: Viruses populate their hosts as a viral quasispecies: a collection of genetically related mutant strains. Viral quasispecies assembly is the reconstruction of strain-specific haplotypes from read data, and predicting their relative abundances within the mix of strains is an important step for various treatment-related reasons. Reference genome independent ('de novo') approaches have yielded benefits over reference-guided approaches, because reference-induced biases can become overwhelming when dealing with divergent strains. While being very accurate, extant de novo methods only yield rather short contigs. The remaining challenge is to reconstruct full-length haplotypes together with their abundances from such contigs. RESULTS: We present Virus-VG as a de novo approach to viral haplotype reconstruction from preassembled contigs. Our method constructs a variation graph from the short input contigs without making use of a reference genome. Then, to obtain paths through the variation graph that reflect the original haplotypes, we solve a minimization problem that yields a selection of maximal-length paths that is, optimal in terms of being compatible with the read coverages computed for the nodes of the variation graph. We output the resulting selection of maximal length paths as the haplotypes, together with their abundances. Benchmarking experiments on challenging simulated and real datasets show significant improvements in assembly contiguity compared to the input contigs, while preserving low error rates compared to the state-of-the-art viral quasispecies assemblers. AVAILABILITY AND IMPLEMENTATION: Virus-VG is freely available at https://bitbucket.org/jbaaijens/virus-vg. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Using the structure of genome data in the design of deep neural networks for predicting amyotrophic lateral sclerosis from genotype.

MOTIVATION: Amyotrophic lateral sclerosis (ALS) is a neurodegenerative disease caused by aberrations in the genome. While several disease-causing variants have been identified, a major part of heritability remains unexplained. ALS is believed to have a complex genetic basis where non-additive combinations of variants constitute disease, which cannot be picked up using the linear models employed in classical genotype-phenotype association studies. Deep learning on the other hand is highly promising for identifying such complex relations. We therefore developed a deep-learning based approach for the classification of ALS patients versus healthy individuals from the Dutch cohort of the Project MinE dataset. Based on recent insight that regulatory regions harbor the majority of disease-associated variants, we employ a two-step approach: first promoter regions that are likely associated to ALS are identified, and second individuals are classified based on their genotype in the selected genomic regions. Both steps employ a deep convolutional neural network. The network architecture accounts for the structure of genome data by applying convolution only to parts of the data where this makes sense from a genomics perspective. RESULTS: Our approach identifies potentially ALS-associated promoter regions, and generally outperforms other classification methods. Test results support the hypothesis that non-additive combinations of variants contribute to ALS. Architectures and protocols developed are tailored toward processing population-scale, whole-genome data. We consider this a relevant first step toward deep learning assisted genotype-phenotype association in whole genome-sized data. AVAILABILITY AND IMPLEMENTATION: Our code will be available on Github, together with a synthetic dataset (https://github.com/byin-cwi/ALS-Deeplearning). The data used in this study is available to bona-fide researchers upon request. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Characteristics of de novo structural changes in the human genome.

Small insertions and deletions (indels) and large structural variations (SVs) are major contributors to human genetic diversity and disease. However, mutation rates and characteristics of de novo indels and SVs in the general population have remained largely unexplored. We report 332 validated de novo structural changes identified in whole genomes of 250 families, including complex indels, retrotransposon insertions, and interchromosomal events. These data indicate a mutation rate of 2.94 indels (1-20 bp) and 0.16 SVs (>20 bp) per generation. De novo structural changes affect on average 4.1 kbp of genomic sequence and 29 coding bases per generation, which is 91 and 52 times more nucleotides than de novo substitutions, respectively. This contrasts with the equal genomic footprint of inherited SVs and substitutions. An excess of structural changes originated on paternal haplotypes. Additionally, we observed a nonuniform distribution of de novo SVs across offspring. These results reveal the importance of different mutational mechanisms to changes in human genome structure across generations.

Genotyping inversions and tandem duplications.

MOTIVATION: Next Generation Sequencing (NGS) has enabled studying structural genomic variants (SVs) such as duplications and inversions in large cohorts. SVs have been shown to play important roles in multiple diseases, including cancer. As costs for NGS continue to decline and variant databases become ever more complete, the relevance of genotyping also SVs from NGS data increases steadily, which is in stark contrast to the lack of tools to do so. RESULTS: We introduce a novel statistical approach, called DIGTYPER (Duplication and Inversion GenoTYPER), which computes genotype likelihoods for a given inversion or duplication and reports the maximum likelihood genotype. In contrast to purely coverage-based approaches, DIGTYPER uses breakpoint-spanning read pairs as well as split alignments for genotyping, enabling typing also of small events. We tested our approach on simulated and on real data and compared the genotype predictions to those made by DELLY, which discovers SVs and computes genotypes, and SVTyper, a genotyping program used to genotype variants detected by LUMPY. DIGTYPER compares favorable especially for duplications (of all lengths) and for shorter inversions (up to 300 bp). In contrast to DELLY, our approach can genotype SVs from data bases without having to rediscover them. AVAILABILITY AND IMPLEMENTATION: https://bitbucket.org/jana_ebler/digtyper.git. CONTACT: t.marschall@mpi-inf.mpg.de. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Repeat- and error-aware comparison of deletions.

MOTIVATION: The number of reported genetic variants is rapidly growing, empowered by ever faster accumulation of next-generation sequencing data. A major issue is comparability. Standards that address the combined problem of inaccurately predicted breakpoints and repeat-induced ambiguities are missing. This decisively lowers the quality of 'consensus' callsets and hampers the removal of duplicate entries in variant databases, which can have deleterious effects in downstream analyses. RESULTS: We introduce a sound framework for comparison of deletions that captures both tool-induced inaccuracies and repeat-induced ambiguities. We present a maximum matching algorithm that outputs virtual duplicates among two sets of predictions/annotations. We demonstrate that our approach is clearly superior over ad hoc criteria, like overlap, and that it can reduce the redundancy among callsets substantially. We also identify large amounts of duplicate entries in the Database of Genomic Variants, which points out the immediate relevance of our approach. AVAILABILITY AND IMPLEMENTATION: Implementation is open source and available from https://bitbucket.org/readdi/readdi CONTACT: roland.wittler@uni-bielefeld.de or t.marschall@mpi-inf.mpg.de SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

SV-AUTOPILOT: optimized, automated construction of structural variation discovery and benchmarking pipelines.

BACKGROUND: Many tools exist to predict structural variants (SVs), utilizing a variety of algorithms. However, they have largely been developed and tested on human germline or somatic (e.g. cancer) variation. It seems appropriate to exploit this wealth of technology available for humans also for other species. Objectives of this work included: a) Creating an automated, standardized pipeline for SV prediction. b) Identifying the best tool(s) for SV prediction through benchmarking. c) Providing a statistically sound method for merging SV calls. RESULTS: The SV-AUTOPILOT meta-tool platform is an automated pipeline for standardization of SV prediction and SV tool development in paired-end next-generation sequencing (NGS) analysis. SV-AUTOPILOT comes in the form of a virtual machine, which includes all datasets, tools and algorithms presented here. The virtual machine easily allows one to add, replace and update genomes, SV callers and post-processing routines and therefore provides an easy, out-of-the-box environment for complex SV discovery tasks. SV-AUTOPILOT was used to make a direct comparison between 7 popular SV tools on the Arabidopsis thaliana genome using the Landsberg (Ler) ecotype as a standardized dataset. Recall and precision measurements suggest that Pindel and Clever were the most adaptable to this dataset across all size ranges while Delly performed well for SVs larger than 250 nucleotides. A novel, statistically-sound merging process, which can control the false discovery rate, reduced the false positive rate on the Arabidopsis benchmark dataset used here by >60%. CONCLUSION: SV-AUTOPILOT provides a meta-tool platform for future SV tool development and the benchmarking of tools on other genomes using a standardized pipeline. It optimizes detection of SVs in non-human genomes using statistically robust merging. The benchmarking in this study has demonstrated the power of 7 different SV tools for analyzing different size classes and types of structural variants. The optional merge feature enriches the call set and reduces false positives providing added benefit to researchers planning to validate SVs. SV-AUTOPILOT is a powerful, new meta-tool for biologists as well as SV tool developers.

Viral quasispecies assembly via maximal clique enumeration.

Virus populations can display high genetic diversity within individual hosts. The intra-host collection of viral haplotypes, called viral quasispecies, is an important determinant of virulence, pathogenesis, and treatment outcome. We present HaploClique, a computational approach to reconstruct the structure of a viral quasispecies from next-generation sequencing data as obtained from bulk sequencing of mixed virus samples. We develop a statistical model for paired-end reads accounting for mutations, insertions, and deletions. Using an iterative maximal clique enumeration approach, read pairs are assembled into haplotypes of increasing length, eventually enabling global haplotype assembly. The performance of our quasispecies assembly method is assessed on simulated data for varying population characteristics and sequencing technology parameters. Owing to its paired-end handling, HaploClique compares favorably to state-of-the-art haplotype inference methods. It can reconstruct error-free full-length haplotypes from low coverage samples and detect large insertions and deletions at low frequencies. We applied HaploClique to sequencing data derived from a clinical hepatitis C virus population of an infected patient and discovered a novel deletion of length 357±167 bp that was validated by two independent long-read sequencing experiments. HaploClique is available at https://github.com/armintoepfer/haploclique. A summary of this paper appears in the proceedings of the RECOMB 2014 conference, April 2-5.

MATE-CLEVER: Mendelian-inheritance-aware discovery and genotyping of midsize and long indels.

MOTIVATION: Accurately predicting and genotyping indels longer than 30 bp has remained a central challenge in next-generation sequencing (NGS) studies. While indels of up to 30 bp are reliably processed by standard read aligners and the Genome Analysis Toolkit (GATK), longer indels have still resisted proper treatment. Also, discovering and genotyping longer indels has become particularly relevant owing to the increasing attention in globally concerted projects. RESULTS: We present MATE-CLEVER (Mendelian-inheritance-AtTEntive CLique-Enumerating Variant findER) as an approach that accurately discovers and genotypes indels longer than 30 bp from contemporary NGS reads with a special focus on family data. For enhanced quality of indel calls in family trios or quartets, MATE-CLEVER integrates statistics that reflect the laws of Mendelian inheritance. MATE-CLEVER's performance rates for indels longer than 30 bp are on a par with those of the GATK for indels shorter than 30 bp, achieving up to 90% precision overall, with >80% of calls correctly typed. In predicting de novo indels longer than 30 bp in family contexts, MATE-CLEVER even raises the standards of the GATK. MATE-CLEVER achieves precision and recall of ∼63% on indels of 30 bp and longer versus 55% in both categories for the GATK on indels of 10-29 bp. A special version of MATE-CLEVER has contributed to indel discovery, in particular for indels of 30-100 bp, the 'NGS twilight zone of indels', in the Genome of the Netherlands Project. AVAILABILITY AND IMPLEMENTATION: http://clever-sv.googlecode.com/

Mapping proteins in the presence of paralogs using units of coevolution.

BACKGROUND: We study the problem of mapping proteins between two protein families in the presence of paralogs. This problem occurs as a difficult subproblem in coevolution-based computational approaches for protein-protein interaction prediction. RESULTS: Similar to prior approaches, our method is based on the idea that coevolution implies equal rates of sequence evolution among the interacting proteins, and we provide a first attempt to quantify this notion in a formal statistical manner. We call the units that are central to this quantification scheme the units of coevolution. A unit consists of two mapped protein pairs and its score quantifies the coevolution of the pairs. This quantification allows us to provide a maximum likelihood formulation of the paralog mapping problem and to cast it into a binary quadratic programming formulation. CONCLUSION: CUPID, our software tool based on a Lagrangian relaxation of this formulation, makes it, for the first time, possible to compute state-of-the-art quality pairings in a few minutes of runtime. In summary, we suggest a novel alternative to the earlier available approaches, which is statistically sound and computationally feasible.

Discovering motifs that induce sequencing errors.

BACKGROUND: Elevated sequencing error rates are the most predominant obstacle in single-nucleotide polymorphism (SNP) detection, which is a major goal in the bulk of current studies using next-generation sequencing (NGS). Beyond routinely handled generic sources of errors, certain base calling errors relate to specific sequence patterns. Statistically principled ways to associate sequence patterns with base calling errors have not been previously described. Extant approaches either incur decisive losses in power, due to relating errors with individual genomic positions rather than motifs, or do not properly distinguish between motif-induced and sequence-unspecific sources of errors. RESULTS: Here, for the first time, we describe a statistically rigorous framework for the discovery of motifs that induce sequencing errors. We apply our method to several datasets from Illumina GA IIx, HiSeq 2000, and MiSeq sequencers. We confirm previously known error-causing sequence contexts and report new more specific ones. CONCLUSIONS: Checking for error-inducing motifs should be included into SNP calling pipelines to avoid false positives. To facilitate filtering of sets of putative SNPs, we provide tracks of error-prone genomic positions (in BED format). AVAILABILITY: http://discovering-cse.googlecode.com.

Mirroring co-evolving trees in the light of their topologies.

MOTIVATION: Determining the interaction partners among protein/domain families poses hard computational problems, in particular in the presence of paralogous proteins. Available approaches aim to identify interaction partners among protein/domain families through maximizing the similarity between trimmed versions of their phylogenetic trees. Since maximization of any natural similarity score is computationally difficult, many approaches employ heuristics to evaluate the distance matrices corresponding to the tree topologies in question. In this article, we devise an efficient deterministic algorithm which directly maximizes the similarity between two leaf labeled trees with edge lengths, obtaining a score-optimal alignment of the two trees in question. RESULTS: Our algorithm is significantly faster than those methods based on distance matrix comparison: 1 min on a single processor versus 730 h on a supercomputer. Furthermore, we outperform the current state-of-the-art exhaustive search approach in terms of precision, while incurring acceptable losses in recall. AVAILABILITY: A C implementation of the method demonstrated in this article is available at http://compbio.cs.sfu.ca/mirrort.htm

CLEVER: clique-enumerating variant finder.

MOTIVATION: Next-generation sequencing techniques have facilitated a large-scale analysis of human genetic variation. Despite the advances in sequencing speed, the computational discovery of structural variants is not yet standard. It is likely that many variants have remained undiscovered in most sequenced individuals. RESULTS: Here, we present a novel internal segment size based approach, which organizes all, including concordant, reads into a read alignment graph, where max-cliques represent maximal contradiction-free groups of alignments. A novel algorithm then enumerates all max-cliques and statistically evaluates them for their potential to reflect insertions or deletions. For the first time in the literature, we compare a large range of state-of-the-art approaches using simulated Illumina reads from a fully annotated genome and present relevant performance statistics. We achieve superior performance, in particular, for deletions or insertions (indels) of length 20-100 nt. This has been previously identified as a remaining major challenge in structural variation discovery, in particular, for insert size based approaches. In this size range, we even outperform split-read aligners. We achieve competitive results also on biological data, where our method is the only one to make a substantial amount of correct predictions, which, additionally, are disjoint from those by split-read aligners. AVAILABILITY: CLEVER is open source (GPL) and available from http://clever-sv.googlecode.com. CONTACT: as@cwi.nl or tm@cwi.nl. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Hybrid-hybrid correction of errors in long reads with HERO.

Although generally superior, hybrid approaches for correcting errors in third-generation sequencing (TGS) reads, using next-generation sequencing (NGS) reads, mistake haplotype-specific variants for errors in polyploid and mixed samples. We suggest HERO, as the first "hybrid-hybrid" approach, to make use of both de Bruijn graphs and overlap graphs for optimal catering to the particular strengths of NGS and TGS reads. Extensive benchmarking experiments demonstrate that HERO improves indel and mismatch error rates by on average 65% (27[Formula: see text]95%) and 20% (4[Formula: see text]61%). Using HERO prior to genome assembly significantly improves the assemblies in the majority of the relevant categories.

Classifying short gene expression time-courses with Bayesian estimation of piecewise constant functions.

MOTIVATION: Analyzing short time-courses is a frequent and relevant problem in molecular biology, as, for example, 90% of gene expression time-course experiments span at most nine time-points. The biological or clinical questions addressed are elucidating gene regulation by identification of co-expressed genes, predicting response to treatment in clinical, trial-like settings or classifying novel toxic compounds based on similarity of gene expression time-courses to those of known toxic compounds. The latter problem is characterized by irregular and infrequent sample times and a total lack of prior assumptions about the incoming query, which comes in stark contrast to clinical settings and requires to implicitly perform a local, gapped alignment of time series. The current state-of-the-art method (SCOW) uses a variant of dynamic time warping and models time series as higher order polynomials (splines). RESULTS: We suggest to model time-courses monitoring response to toxins by piecewise constant functions, which are modeled as left-right Hidden Markov Models. A Bayesian approach to parameter estimation and inference helps to cope with the short, but highly multivariate time-courses. We improve prediction accuracy by 7% and 4%, respectively, when classifying toxicology and stress response data. We also reduce running times by at least a factor of 140; note that reasonable running times are crucial when classifying response to toxins. In conclusion, we have demonstrated that appropriate reduction of model complexity can result in substantial improvements both in classification performance and running time. AVAILABILITY: A Python package implementing the methods described is freely available under the GPL from http://bioinformatics.rutgers.edu/Software/MVQueries/.

Determining coding CpG islands by identifying regions significant for pattern statistics on Markov chains.

Recent experimental and computational work confirms that CpGs can be unmethylated inside coding exons, thereby showing that codons may be subjected to both genomic and epigenomic constraint. It is therefore of interest to identify coding CpG islands (CCGIs) that are regions inside exons enriched for CpGs. The difficulty in identifying such islands is that coding exons exhibit sequence biases determined by codon usage and constraints that must be taken into account. We present a method for finding CCGIs that showcases a novel approach we have developed for identifying regions of interest that are significant (with respect to a Markov chain) for the counts of any pattern. Our method begins with the exact computation of tail probabilities for the number of CpGs in all regions contained in coding exons, and then applies a greedy algorithm for selecting islands from among the regions. We show that the greedy algorithm provably optimizes a biologically motivated criterion for selecting islands while controlling the false discovery rate. We applied this approach to the human genome (hg18) and annotated CpG islands in coding exons. The statistical criterion we apply to evaluating islands reduces the number of false positives in existing annotations, while our approach to defining islands reveals significant numbers of undiscovered CCGIs in coding exons. Many of these appear to be examples of functional epigenetic specialization in coding exons.